All in all, our results support feature distortion + grammaticality bias accounts, oppose inhibitory similarity-based interference, and suggest that at least for some online processing effects, task plays a much more important role than we used to think.

Changing the expected task radically changed the effects in reading times. But why? Christopher Hammerly proposed that grammaticality judgement task implies that some sentences will be ungrammatical, which shifts participants' grammaticality bias and allows attraction effect to surface

Turns out, all previous studies used comprehension questions, while our participants expected to judge sentence grammaticality. In a follow-up experiment, we changed the expected task to questions, and found no trace of agreement attraction.